Block has introduced Dessa, a speech-to-speech AI model developed by its internal research lab. The system is trained exclusively on human conversation, omitting all text-based data, books, or web pages during the pretraining process. The model aims to replicate natural human interaction by learning language through listening.
The model operates in full-duplex mode, allowing it to listen and speak simultaneously. This architecture enables the AI to manage interruptions, emotional tone, and conversational rhythm. During demonstrations, the model showcased its ability to handle pauses and laughter, though it occasionally exhibited errors, such as misidentifying a user.
Dessa is currently in early access, with priority access granted to AI safety researchers. Block intends to evaluate the model's risks and implications regarding trust and impersonation before considering a broader release. The project is led by Ragavan Thurairatnam, who previously co-founded the startup Dessa, which Block acquired in 2020.
This initiative represents a shift in Block's AI strategy, moving beyond its previous open-source AI agent project codenamed goose. By focusing on speech-native intelligence, the company aims to refine how users interact with AI systems. The development team is currently gathering external feedback to analyze the behavior of the model.
